Hongjie Jiang

Ionically conductive hydrogels are promising materials for flexible sensors, electronic skin, and wearable electronics owing to their compliance, stretchability, and ionic conductivity. However, the formulation design of multicomponent hydrogels still largely relies on trial-and-error experiments, making it inefficient to simultaneously optimize multiple electromechanical properties. Here, we propose a data-driven inverse design strategy for ionically conductive hydrogel formulations. Based on 233 experimentally measured formulation–property samples, a conditional formulation generation model based on a conditional Wasserstein generative adversarial network with gradient penalty was first developed to generate candidate formulations under target property constraints. A component-aware hydrogel graph Transformer model, HydroGT, was then constructed to represent each formulation as a component graph for multiproperty prediction and candidate screening. HydroGT achieved an average R2 of 90.44% and an average NRMSE of 6.76% for five-property prediction, outperforming conventional regression models. Experimental validation showed that the formulation obtained by the complete generation–screening framework exhibited lower relative errors across all five target properties than the formulation generated without screening. This work provides an efficient data-driven strategy for the target-oriented design of multicomponent ionically conductive hydrogels.

Flexible thin-film pressure sensors are widely used in tactile sensing applications, yet their performance is often position-dependent, compromising measurement accuracy and repeatability. In this study, a pressure-concentrating structure was developed to guide loads from different contact points to a fixed sensitive region, significantly improving response consistency and low-force detection. The optimized sensors were integrated into a dexterous-hand system for pressure-based weight estimation, while a MobileNetV1-based vision module was used for camera-based object classification. Additionally, a smart cushion equipped with a 4 × 4 sensor array was designed for sitting-posture recognition, with Gaussian-kernel-based reconstruction enabling continuous pressure mapping. Machine learning models, including SVM and random forest, demonstrated high accuracy in both object and posture recognition tasks. The proposed system reduces position-dependent response variation and improves response consistency within the tested force range, enabling reliable multimodal perception and practical applicability in real-time grasping and posture-monitoring scenarios.

Echinoderms, such as sea cucumber, dynamically change their body size to environmental stimuli. By contrast, typical synthetic materials cannot reinvent their structures once formed. We propose a strategy for developing “self‐extendability” polymeric materials that can be structurally tuned by repetitive mechanical stress stimulation. Polyion complex glycerol (PICG) hydrogels undergo self‐extendability and mechanical consistency, and the materials substantially grow in deformation under repetitive loading through a diffusion of disentangled dense structure to sparse structure. This strategy is generalizable to other polymers and topologies. Additionally, the gel is able to adhere to 3D surface. These advancements make our gels ideal for e‐skin substrate materials; they simultaneously offer extended performance and curved surface attachment. This work may open an avenue for the development of self‐extendability materials for intelligent devices.

Carbon black (CB) is attractive for electrothermal films owing to its low cost and facile processability, but agglomeration and weak filler‐matrix interactions often undermine conductivity, adhesion, and heating uniformity. Here, we regulate the CB‐polyurethane (PU) interface using 3‐Glycidyloxypropyltrimethoxysilane (GPTMS) and elucidate how the silane dosage governs CB dispersion, percolation behavior, and the Joule‐heating performance of CB/PU films. GPTMS grafts onto CB via hydrolysis to form Si‐O‐C/Si‐O‐Si interfacial linkages, while the epoxy group undergoes ring‐opening reactions with PU, thereby enhancing CB‐PU coupling. Moderate GPTMS loadings (GPTMS:CB = 1:1–3:1) yield well‐dispersed, continuous conductive networks in PU, producing low resistivity, strong adhesion, and substantially improved thermal uniformity. In contrast, excessive GPTMS (≥ 4:1) promotes siloxane self‐condensation, reinstates CB aggregation, and introduces insulating interlayers that disrupt electron pathways. Interestingly, the CB volume fraction‐dependent conductivity shows that silane modification increases the percolation threshold by thickening interfacial layers (tunneling barrier) yet steepens the post‐threshold conductivity rise due to more effective cluster growth. Printed films (with PU:CB = 5:1, GPTMS:CB = 3:1) demonstrate application‐level performance that under 5 V, they rapidly reach ~55°C within 15 s with uniform temperature fields and retain both heating and adhesion after repeated bending. This study establishes a mechanistic framework linking silane‐mediated interfacial chemistry to network topology and device metrics and provides a practical processing window (GPTMS:CB ≈ 2–3:1) for scalable, durable CB‐based electrothermal films and wearable heaters.

Optimizing multimodal models using unimodal objectives has proven effective in mitigating modality bias and enhancing generalization. This multitask-style training paradigm has motivated a range of gradient-based optimization methods. However, existing approaches predominantly focus on adjusting gradient magnitudes during aggregation, while overlooking the directional inconsistency between the final update direction and the gradients derived from unimodal and multimodal objectives, which may result in suboptimal optimization. To address this issue, we propose Scalable Decomposition Gradients (SDG), a novel framework for conflict-free optimization in multimodal learning. SDG explicitly decouples the optimization of gradient magnitude and direction. Specifically, we employ modality-aware orthogonal projections to remove conflicting components across gradients, ensuring that the final update direction aligns with both unimodal and multimodal learning objectives. We further utilize the projection residuals to adaptively scale each gradient, balancing their relative contributions and mitigating magnitude disparities. Theoretical analysis demonstrates that such conflict-free updates with consistent step sizes promote balanced convergence across losses from different modalities. Extensive experiments on several widely used multimodal benchmarks validate the effectiveness of SDG, showing notable performance improvements without introducing additional model complexity.

Recent advances in electrochemical capsaicin detection, highlighting progress in electrode sensing materials and their translation into sensing technologies.

Sign language recognition technology holds significant importance for eliminating communication barriers faced by the hearing‐impaired population. To address the limitations of current wearable sensors‒such as complex fabrication or materials incompatibility within an integrated system, this study designed a 3D printed ion‐conductive hydrogel with tunable electromechanical performances for versatile wearable sensing. The hydrogel is primarily based on a polyampholyte network interpenetrated with a polyacrylamide (PAAM) framework and synergistically integrated with LiCl and a covalent organic framework (COF) to enhance its electromechanical performance. It exhibits low hysteresis (90.25% recovery ratio) with high elongation (550%) and large compressive strain tolerance (90%) for strain/pressure sensing, while its low modulus (0.09 MPa) and high conductivity (0.23 S m ^{− 1} ) enabled high‐fidelity surface electromyography (sEMG) sensing. Leveraging these multifunctional hydrogels, we developed a multimodal sign language recognition system consisting of a pair of digital gloves, each embedded with 12 strain sensors and 5 pressure sensors, together with a flexible armband integrated with a 10‐channel differential sEMG electrode array. Coupled with a bidirectional long short‐term memory (Bi‐LSTM) multimodal fusion model, the system achieved a classification accuracy of 99.65% across 24 Chinese sign language gestures.